feat(ai): external responder for external-* model ids (reference runs)

모델 id가 external- 로 시작하면 ai-agent 가 LLM API 를 부르지 않고 각 턴(시스템
프롬프트·대화 이력·도구 정의·강제 제출 여부)을 GET /external/next 로 내어주고
POST /external/reply 로 답을 받는다. 진단 루프·프롬프트·도구·턴 상한은 다른 모델과
동일하므로, API 경로가 없는 모델(Claude Mythos 5.1 을 Claude Code 서브에이전트가
구동)의 참고 측정에 쓴다. 토큰 사용량은 알 수 없어 0 으로 기록한다.

external-* 진단은 EXTERNAL_DIAGNOSE_TIMEOUT_S(기본 1500초), 한 턴은
EXTERNAL_TURN_TIMEOUT_S(기본 1200초)까지 기다리고, 하네스는 같은 모델에 26분을
기다린다. /selfcheck 에 external_responder 상태(대기 턴·마지막 폴링 시각)를 추가.

ai/internal/agent/agent.go

@@ -45,6 +45,7 @@ type Agent struct {
 	cfg        Config
 	oai        *openai.Client
 	anth       Responder // nil = claude-* models unavailable
+	ext        Responder // nil = external-* models unavailable (internal/external: a client outside the VM answers each turn)
 	tools      *tools.Runner
 	harness    *harness.Client
 	store      *rag.Store
@@ -74,7 +75,14 @@ func New(cfg Config, oai *openai.Client, tr *tools.Runner, hc *harness.Client, s
 // SetAnthropic enables claude-* model ids (Claude Messages API adapter).
 func (a *Agent) SetAnthropic(c Responder) { a.anth = c }
 
-// provider picks the LLM client for a model id: claude-* -> Anthropic adapter, everything else -> OpenAI.
+// SetExternal enables external-* model ids (turns served to an outside client over /external/next, reference runs).
+func (a *Agent) SetExternal(c Responder) { a.ext = c }
+
+// IsExternal reports whether the model id is served by the external responder (longer deadlines apply).
+func IsExternal(model string) bool { return strings.HasPrefix(model, "external-") }
+
+// provider picks the LLM client for a model id: claude-* -> Anthropic adapter, external-* -> external responder,
+// everything else -> OpenAI.
 func (a *Agent) provider(model string) (Responder, error) {
 	if strings.HasPrefix(model, "claude-") {
 		if a.anth == nil {
@@ -82,6 +90,12 @@ func (a *Agent) provider(model string) (Responder, error) {
 		}
 		return a.anth, nil
 	}
+	if IsExternal(model) {
+		if a.ext == nil {
+			return nil, fmt.Errorf("model %s: external responder not configured", model)
+		}
+		return a.ext, nil
+	}
 	return a.oai, nil
 }
 

ai/internal/agent/agent_test.go

@@ -127,3 +127,25 @@ func TestDiagnoseRoutesClaudeModel(t *testing.T) {
 		t.Fatalf("usage: %v", res.Usage)
 	}
 }
+
+// Model routing: external-* goes to the external responder when one is set, and is refused when none is configured.
+func TestProviderRouting(t *testing.T) {
+	a := New(Config{}, nil, nil, nil, nil, diag.Enums{}, "v")
+	if _, err := a.provider("external-claude-mythos-5-1"); err == nil || !strings.Contains(err.Error(), "external responder not configured") {
+		t.Fatalf("expected unconfigured error, got %v", err)
+	}
+	fake := &fakeResponder{}
+	a.SetExternal(fake)
+	if p, err := a.provider("external-claude-mythos-5-1"); err != nil || p != Responder(fake) {
+		t.Fatalf("external routing: %v %v", p, err)
+	}
+	if !IsExternal("external-x") || IsExternal("claude-x") || IsExternal("gpt-6-luna") {
+		t.Fatal("IsExternal prefix check")
+	}
+}
+
+type fakeResponder struct{}
+
+func (*fakeResponder) CreateResponse(context.Context, openai.Request) (*openai.Response, error) {
+	return nil, nil
+}